lessons-reviewlisted

# Lessons Review `.claude/.harness/candidates.jsonl` is the **episodic** layer: raw incidents captured automatically and never injected into context. This skill is the distillation pass that turns a few of them into semantic memory and throws the rest away. **Throwing most of them away is the expected outcome.** Capture is generous on purpose; strictness lives here. ## Read the backlog ```bash node .claude/harness/candidates.mjs list ``` Read all of it before deciding anything. Patterns across entries are the point — three separate corrections about the same thing are one lesson, and you can only see that having read all three. ## For each candidate, choose one of four **1. Encode it as a guard.** The best outcome, and the most often skipped. If the incident is mechanically checkable — a command that should never run, a file that should never be written, a shape that is always wrong — write the check instead of the lesson. A guard fires every time; a lesson only fires when its triggers happen to match. Add it to `.claude/harness/`, wire it in `settings.json`, and **add both-direction test cases**. The allow cases matter more than the block cases: a guard that refuses ordinary work gets switched off. **2. Promote it to `CONVENTIONS.md`.** If it is a standing fact about the project rather than a behaviour change — where something lives, what the registration checklist is — it belongs there, where it is read on every planning task rather than only on a keyword match. *